# Source and modification notices

This exploratory reference uses selected images and publisher-generated masks
from **Controlled Diet (CD) Dataset for Methane Plume Detection**, by Mohamed G.
Embaby, Toqi Tahamid Sarker, Amer AbuGhazaleh and Khaled R. Ahmed, Southern
Illinois University Carbondale.

- Dataset: [baselab/controlled-diet-methane-dataset](https://huggingface.co/datasets/baselab/controlled-diet-methane-dataset).
- Exact revision: `9475f96c32dd4764471e4358ca6659b0dd348ecd`.
- [Pinned LICENSE](https://huggingface.co/datasets/baselab/controlled-diet-methane-dataset/blob/9475f96c32dd4764471e4358ca6659b0dd348ecd/LICENSE)
  and [dataset card](https://huggingface.co/datasets/baselab/controlled-diet-methane-dataset/blob/9475f96c32dd4764471e4358ca6659b0dd348ecd/README.md)
  explicitly declare **Creative Commons Attribution 4.0 International**.
- [License terms](https://creativecommons.org/licenses/by/4.0/).
- Associated paper: [Optical gas imaging and deep learning for quantifying enteric
  methane emissions from rumen fermentation in vitro](https://doi.org/10.1049/ipr2.13327),
  IET Image Processing, 2025. This pilot is not a reproduction of its models or
  chemical-measurement result.

The source metadata JSON separately declares CC0-1.0. That inconsistency remains
unresolved. This work elects the explicit CC BY 4.0 grant, keeps attribution and
change notices, and makes no public-domain claim. No author endorsement is implied.
Source PNG bytes are preserved; their full hashes are the pinned LFS SHA256.

Changes for this work: a prospectively chosen compact subset, entire recording
IDs assigned to disjoint development roles, classes collapsed to nonzero binary
mask occupancy, 320×240 rendered RGB preprocessing and fractional 2×2 training
targets, native-resolution scoring, conventional and compact from-scratch CPU
references. The original filenames, bytes and publisher mask codes remain in the
source manifest. The code 2/class 3 and code 3/class 2 correspondence is checked.

The new original implementation is MIT licensed. NumPy and Pillow retain their
own notices in the pinned official dependency wheels. No pretrained gas weights,
Gasformer/LangGas author software, client architecture, plant telemetry or robot
capture is included. Learned scores reproduce weak mask labels; they are not
physical methane confidence, concentration, leak absence or field qualification.